OpenAI has long held the belief that they have a unique position in the market and therefore a very wide and deep moat protecting them from any competition coming along and eating their lunch. That moat, seemingly, is due only to the massive and ever-increasing cost of training one of their AI models. Their belief, then, was that they were so far ahead of everyone else and had so much more capital available to them that they could withstand any onslaught and simply outspend the competition. They could therefore justify any capital expenditure as a means to keep the moat filled and sustain their market “leadership”.
That’s all changed in the past few weeks.
What OpenAI and the rest of the market seemingly didn’t foresee was the fact that you don’t really need to spend $2+ billion to train a model - you can train one for about $6 million over the span of just a couple months. The latest model (R1) from Chinese AI startup DeepSeek was trained in a very short time using open datasets, and released for free to anyone who wants it. Benchmarks of these models are difficult to trust fully but initial user testing seems to indicate that it outperforms OpenAI’s o1 models in most metrics and blows away similar offerings from Anthropic.
DeepSeek-R1 spells the end of OpenAI’s market dominance. A small startup in China was able to in the span of a few months create a new model that matches or outperforms o1 using 2+ year old hardware (Nvidia H800, which is a China-specific model of the 3 year old H100 chip) for only $6 million in energy and hardware costs. Oh, and anyone with an Apple Silicon Mac can run it at home, for free, with a few minutes of setup.
OpenAI meanwhile is investing $500 billion into building new datacenters in hopes that they can somehow maintain their dominance that way. I just don’t see it. I’ve been saying for a long time that there’s nothing special about o1 or Claude or Gemini or Llama. They’re trained on stolen and scraped data from the internet using techniques that have been known to AI professionals for over a decade at this point. Sam Altman has conned the market into believing that the only way to make these models is for investors to give him personally a trillion dollars for him to make bigger and bigger slop machines. Never mind innovation! Just throw more and more money onto the fire to produce mediocre results year after year.
That ends now. DeepSeek proves that the OpenAI way of doing things is a sham. You don’t need the latest hardware to train these models. You don’t need a $500 billion datacenter or your own chip fab. You just need a bunch of GPUs and some dedicated engineers. Anyone can pull this off, and I hope that more and more companies from large enterprise to startup pull the wool off their eyes and realize that they could make their own models (or just use DeepSeek) for the equivalent cost of 2 months of OpenAI API credits.
Keep holding on. The bubble is nearly over.
Jan 27, 2025 Update
